Conceptually, a smoother is a function from $x$ to $y$, which is defined at every $x$ value (or at least, every $x$ value close enough to a data point). There are infinitely many possible $x$ values, so that's not practical.
If you want the smoother in order to draw it on the scatterplot (or draw it instead of the scatterplot), you need it for at most as many $x$ values as you have pixels, and you can often make do with many fewer (maybe 100 or so) with linear interpolation done by the graph drawing process. The number of points you need and their spacing doesn't really depend on what data you were given, so in that sense it's unrelated to the values of the $x_i$.
More specifically:
- you don't need to predict at every $x_i$, especially if there are lot of them
- you may need to predict at some values $x$ not in $x_i$, especially if there are gaps in the distribution of the $x_i$.
It is common for smoothers to specify the vectors $\langle x_i, y_i\rangle$ as inputs and another vector called something like xout whether the smoother is to be evaluated. It is less universal, but not uncommon, for the default xout to be a uniformly spaced grid.
As a secondary note, I don't think the claim is strictly true. First, there are settings where an analytic formula for the smoother is a feasible output (eg, penalised splines) and where that is valuable to simplify taking derivatives and integrals. Second, it's not hard to find implementations where the default behaviour is to predict at the input values $x_i$, eg
xoutfor smoother to be drawn on a graph or make predictions. What I am still confusing is why thesexoutshould "not related to $x_i$'s", like what my lecture note says. My understanding is even if $x$ is to some extent related to the $x_i$'s, we still can draw reasonable smoothing curves. – WCMC Sep 07 '20 at 15:57